feat(client): compose the production HarnessClient from a profile name

[chughtapan]

Plan step 9. Stacked on #954. This is the step that makes a production HarnessClient constructible at all.

What was missing

The only code that started a daemon and connected a client to it lived in test-utils, demanded a caller-supplied port, and needed an injected checkpoint store. Combined with the slot change in #954, nothing was left blocking — but nothing was wired either.

The composition

harnessClientForProfile(profileName)   // that is the entire input

It starts the slot's own daemon child, derives the loopback endpoint from the same persisted port the daemon binds, and provides a file-backed checkpoint store. No URL, no port, no store from the caller — which is exactly what the adapters need in steps 10 and 11.

The checkpoint directory is keyed by profile name, not AgentId. That is forced, not stylistic: acquireHarnessClient reads agentId from the status tool inside its own body, so the identity does not exist at the moment the store must be provided. One slot is exactly one AgentId, so the profile name is a valid agent scope.

The daemon supervisor moves from test-utils/process/packaged-moltzapd.ts to src/moltzapd-child.ts via git mv, so the proven readiness-polling and shutdown behavior is preserved rather than rewritten. Only the port reservation a test performs before writing a slot stays behind, in test-utils/process/reserve-port.ts.

The round-trip test now proves production

harness-adapters.integration.test.ts previously called the test-only acquirer and supplied KeyValueStore.layerMemory. It now calls harnessClientForProfile — so the one ungated end-to-end proof for both adapters exercises the real composition and a real file-backed store. acquireHarnessClient and KeyValueStore became unused imports there, which is the clean signal that the parallel test path is gone.

Architecture findings this surfaced, and how they were resolved

Three real ones, none suppressed blindly:

1. require-boundary-owned-types — harness-client.ts may not mention node directly. Resolved by moving the node-touching composition into the node-boundary module rather than waiving the rule.
2. Folder dependency cycle — my first fix put the module under harness/, creating . ↔ harness. Reverted; the module sits at the source root where its imports point one way.
3. no-large-folder / file-implicit-boundary-module — the package genuinely gained a module, and harness-client.ts genuinely is a boundary (it is the ./harness-client subpath export). Both recorded in scripts/gen-architecture-configs.mjs, since the config is generated and a hand edit would be overwritten by arch:config:check.

connectMcpOnce was also rewritten Effect-natively instead of carrying an async suppression into production source — sloppy-code-guard and the eslint rule were both asking for the same thing, and the rewrite drops a try/catch too.

Gates

• pnpm nx run workspace:precommit — exit 0 (this runs lint:sloppy-code-guard, which pnpm lint alone does not)
• pnpm typecheck and pnpm nx run-many -t typecheck:tests — 0 errors
• pnpm lint — 0 warnings, 0 errors; arch:check 0 findings on every package
• pnpm nx run-many -t test — all 9 projects green
• git grep layerMemory packages/ returns only pure-unit tests

ADR conformance (R-pass)

Governing outcome	Record	Owner	Binds this diff?	Verdict	Evidence
Daemon and adapter construct http://127.0.0.1:<mcpPort>/mcp; port-zero and bind fallback rejected	20260728-endpoint-daemon-speaks-modern-mcp	v2	no — followed anyway	CONFORMS	The URL is derived from the slot; nothing allocates, discovers, or falls back. A port already in use surfaces the daemon's existing bind failure.
Checkpoints are client-owned presentation state	20260801-harness-client-owns-runtime-context	v2 (contested)	no	CONFORMS	The store is provided at the composition site; the client keeps ownership. No fsync, sharding, cache-layout, or corruption policy added — layerFileSystem stock behavior is accepted as-is.
—	—	main	—	—	No main-governing record covers daemon supervision or checkpoint storage. This composition implements the shape the maintainer selected; the main-owned ADR bundle (step 5) is still outstanding.

No new divergence introduced; no #926 ledger row changes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01P76aaa1STr3WPZ3nDascta

[chughtapan]

Superseded by #974, which collapses this stack into one PR against main. This branch's commits are all contained in it — nothing is lost, and this thread stays readable. Closing to leave one place to review.